A novel feature reduction framework for reliable delta-radiomics of prostate cancer response to radiation therapy

Motivation: Early detection of recurrent prostate cancer (PCa) following primary radiation therapy (RT) can enable salvage treatments. Delta-radiomics from longitudinal quantitative MRI shows promise as treatment response imaging biomarkers. Goal(s): To develop a framework for identifying reliable radiomic features for robust, early detection of response to RT in PCa. Approach: Voxel-wise radiomic feature maps were extracted from longitudinal quantitative MRI and assessed for their ability to detect treatment-related changes, accounting for measurement uncertainties. Results: Only 20% of features with significant post-RT changes were reliably detectable. This framework can identify features that may improve predictive models for treatment response monitoring. Impact: A framework for identifying reliable features in delta-radiomics of longitudinal prostate qMRI was developed. Predictive models built with these robust radiomic features may enable early detection and localisation of recurrences, providing an opportunity for salvage therapies of treatment-resistant tumours.